CN103488485A - Parallel migration method for optimizing operating performance of application software on server - Google Patents

Abstract

The invention provides a parallel migration method for optimizing the operating performance of application software on a server. The parallel migration method re-establishes and migrates density functional energy terms and optimizes and analyzes a density functional theory in a parallel migration mode so as to achieve the purposes of software optimization and quantification. By adopting the parallel migration method, the space of the application field can be effectively widened, the limitation of a single optimizer is avoided, and accordingly resources of clusters and a super computer are effectively utilized. The parallel migration method has considerable effectiveness on the simple optimization of library functions and instruction sets. By means of the parallel migration method, the availability of the density functional theory on computational chemistry, material science and nanoscience is accelerated, the density functional theory can be conveniently and effectively optimized and analyzed, and the study efficiency in the fields of the computational chemistry, the material science, the nanoscience and the like can be improved.

Description

A Parallel Migration Method for Optimizing the Running Performance of Application Software on Server

technical field

The present invention relates to the application of high-performance computing in the fields of computational chemistry, material science and nanotechnology. Specifically, a method of parallel transplantation is used to reorganize and transplant density functional energy items, and to optimize and analyze density functional theory. To achieve the purpose of optimizing the quantification software. the

Background technique

With the continuous advancement of science, people's cognition of matter has extended to the level of electrons. The establishment of the Schrödinger equation makes the theoretical framework of the material world fully established. Through the means of quantum mechanics, people have been able to accurately grasp the shape of objects at the electronic, atomic or molecular scale. But the difficulty in the middle is how to solve the Schrödinger equation that goes deep into the electronic structure. So far, there are mainly two types of computational theories about electronic structure, one is the ab initio quantitative theory based on Hartley Falk (HF) theory or other theoretical models based on HF. Energy and other properties of matter can be obtained with arbitrary precision through the calculation of these theories. In addition, Hartree-Fock (HF) and other theories related to it can systematically correct the calculation accuracy, but the calculation amount of this theoretical method is particularly large, which will consume huge computing resources, and the outer electrons can only Consider 4-fold orbitals of several atoms. Another method for calculating electronic structure is density functional theory (DFT). Its main advantage over HF theory is that the calculated system can be expanded to a system with a particle number of N ³ . In DFT theory, energy is separable. Its energies include non-interacting kinetic energy, electron-nucleus interaction energy, electron-electron interaction potential energy, and exchange-correlation energy. The current DFT theory mainly uses "Jacob's Ladder Technique" to study better and more detailed descriptions of exchange-correlation functional terms.

The first three energies in density functional theory can be used to analyze and update the exchange correlation term, and the calculation of the exchange correlation term and the differential of the correlation term must rely on a series of system grids. Even if the calculation of the exchange-correlation energy formed by the density can only be performed on the system grid points, a series of system grids certainly do not have arbitrary precision analysis and calculation, but the calculation depends on the complex system grid. Moreover, a series of energy calculations on the system grid during the calculation depends on the orientation of the molecules relative to the system grid, which obviously does not have any physical meaning. In addition, the re-improvement of the system grid can be realized, which can shorten the calculation speed. the

In order to improve the operation speed of DFT theory software on the server, a parallel transplantation method is designed here to optimize and analyze the DFT software program, and achieved ideal results. This parallel porting method has many advantages, the main advantage of which is scalability, which also breaks that a pure optimizer may be trapped in a very small application range and cannot find a better available point. The parallel transplantation method here broadens the application field for the simple transplantation method, and it is easier to find the parameters of the DFT software that can be optimized , to speed up the speed of software convergence.

Contents of the invention

The purpose of the present invention is to provide a parallel transplantation method for optimizing the running performance of application software on a server. the

The object of the present invention is achieved in the following manner, and the optimized parameter formula is as follows,

=

，得到其它的相对平均误差

，该数据的确定对后面的实现平行移植方法有决定意义，能够实现单移植方法的快速收敛，并对计算的结果无影响，基本的程序架构和组织流程为： corrected optimal absolute mean error

, to get other relative mean errors

, the determination of this data has decisive significance for the implementation of the parallel transplantation method later, and can realize the rapid convergence of the single transplantation method without affecting the calculation results. The basic program structure and organization flow are:

； 1) Test one by one the optimization parameters of each element in the DFT theory software corresponding to the single transplantation method

;

,测试并找到对应每种元素的最优参数值； 2) 0.5

, test and find the optimal parameters for each element value;

的绝对平均误差的导数

； 3) Calculation of atomic energy

The derivative of the absolute mean error of

;

，可以得到其它的相对平均误差

，这些数值即为并行移植方法中的初始优化参数； 4) Correct the optimal absolute mean error

, other relative average errors can be obtained

, these values are the initial optimization parameters in the parallel transplantation method;

5) Install the parallel version of the transplant method program Python;

6) Realize the combination of DFT application software and transplantation method program Python;

。 7) Find the optimal optimization setting parameters for each element

.

The beneficial effect of the present invention is : the present invention aims at some new technical improvements and inventions for the limited application field of the single transplant method and the low operating performance of the application software on the server. This technical invention is mainly on the basis that the single-transplant method can well optimize the application software of DFT theory, and adopts a parallel method to turn the single-transplant optimizer into a parallel optimizer, which greatly improves its application range and reaches Effectively optimize the performance of application software running on the server.

Detailed ways

The invention describes a method of using parallel transplantation to accelerate and optimize the usability of density functional theory. In the process, the data communication volume when the quantitative software is running on the server can be effectively reduced, so that multiple cores on multiple cluster nodes can be more effectively utilized. the

In order to make the purpose, technical solution and advantages of the present invention clearer, we use the guassian application software as an example to describe the key steps in the present invention in detail, and provide the parameters for realizing setting optimization in detail. the

The optimization parameter formula is as follows,

=

=

，可以得到其它的相对平均误差

。该数据的确定对后面的实现平行移植方法有决定意义，能够实现单移植方法的快速收敛，并对计算的结果无影响。 corrected optimal absolute mean error

, other relative average errors can be obtained

. The determination of this data has decisive significance for the implementation of the parallel transplantation method later, and can realize the rapid convergence of the single transplantation method without affecting the calculation result.

The basic program structure and organizational process are:

； 1. Test one by one the optimization parameters of each element in the DFT theoretical software corresponding to the single transplantation method

;

,测试并找到对应每种元素的最优参数

值； 2.0.5

, test and find the optimal parameters for each element

value;

的绝对平均误差的导数

； 3. Calculation of atomic energy

The derivative of the absolute mean error of

;

，可以得到其它的相对平均误差，这些数值即为并行移植方法中的初始优化参数； 4. Correct the optimal absolute mean error

, other relative average errors can be obtained , these values are the initial optimization parameters in the parallel transplantation method;

5. Install the parallel version of the transplant method program Python;

6. Realize the combination of DFT application software and transplantation method program Python;

7. Find the optimal optimization setting parameters for each element

The parallel transplantation method adopted in the present invention can greatly accelerate the optimization of the DFT application software. For the use of the Guassian application software as an example, its optimization parameters are found, it is relatively easy to realize its optimization settings, and the limited computing resources can be greatly utilized. Accelerate the scientific research of computational chemistry, material science and nanoscience, and can reduce the loss of the server and improve the computing performance of the server, and it is more energy-saving. Application example Gaussian application software optimization parameter value of the present invention is as follows: Element PGA h 0.758 Li 0.565 be 0.879 B 0.617 C 0.667 N 0.665 o 0.645 f 0.613 Na 0.545 Al 0.754 Si 0.523 P 0.813 S 0.705

The invention fully analyzes the theoretical operation characteristics of DFT application software, fully utilizes the program Python that can realize the parallel transplantation method, and greatly accelerates the operation speed of the application software on the server. This method is an effective way to solve the time-consuming operation of application software on the server. It has been widely used and promoted, and it greatly facilitates researchers to do more in-depth exploration and discovery of micro and nano systems.

Except for the technical features described in the instructions, all are known technologies by those skilled in the art. the

Claims (1)

1. a parallel implantation method of optimizing application software runnability on server, is characterized in that the Optimal Parameters formula is as follows,

=

Proofread and correct optimum absolute average error , obtain other relative average error

, these data determine to back realize that the capable of parallel moving method for planting is marginal, can realize the Fast Convergent of single implantation method, and on the result calculated without impact, basic program architecture and organization flow are:

1) test one by one the Optimal Parameters of every kind of element in the theoretical software of the corresponding DFT of single implantation method ;

2) 0.5

, test and find the optimized parameter of corresponding every kind of element

value;

3) calculate atomic

the derivative of absolute average error

;

4) proofread and correct optimum absolute average error

, can obtain other relative average error

, these numerical value are the initial optimization parameter in parallel implantation method;

5) the implantation method program Python of parallel version is installed;

6) realize the combination of DFT application software and implantation method program Python;

7) find the optimization parameters of the optimum of each element

.